← BackReference (opens in a new tab)

The approval conveyor belt

Human review exists but reviewers cannot assess the action. · PM Drills · 27 · 30 sec–2 min

The approval conveyor belt · 30 sec–2 min

Situation

Reviewers approve almost every AI recommendation.

They see only a summary and have seconds per case. Critical source evidence is hidden several clicks away.

Think before scrolling

Does this review meaningfully reduce risk?

What would you change and measure before relying on it?

How a strong PM thinks

Make the important error detectable.

Show the current state, proposed change, consequence, and relevant evidence together. Give reviewers a clear rubric, adequate time, and the ability to reject or escalate.

Sample approved cases to estimate missed errors. Measure review effort and queue age. If meaningful checking is infeasible, constrain the automation's scope rather than treating an approval click as protection.

Takeaway

Review is a capability, not a checkbox.

A human must be able to detect the relevant failure before the consequential action occurs.